rationality of strategies in games theory · rationality of strategies in games theory shichang...
TRANSCRIPT
Rationality of Strategies in Games Theory Shichang Zhang
Stat 157
Professor Aldous
April 20 2016
Abstract
This is an extended write-up of analyzing a well-known game interested both psychologists and
game theorists called “guess 2/3 of the average”; all the players choose an integer from 1-100
simultaneously; whose choice is the closest to 2/3 of the average of all these numbers will win and
get 50 dollars as a prize. This paper also analyses two variations of the original game. One variation
changes the way of calculating the winning number, and the other changes the prize amount. The
analysis focuses on two perspectives. One is breakdown the game and analyze different types of
strategies on the psychological side to illustrate different levels of rationality of these strategies.
The other one is trying to come up with an optimal strategy that will win the game. I build a simple
model for the optimal strategy using it to simulate data and compare the result with a larger dataset
from the previous experiment with New York Times readers.
1. Introduction
The game “Guess 2/3 of the Average” was created by Alain Ledoux in 1981. He used this
game as a tie breaker to select the winner among 4000 magazine readers, who reached the same
amount of points in the previous puzzle. Then Rosemarie Nagel revealed that this kind of games
can be used to disclose participants’ “depth of reasoning”. Also, due to the analogy of Keynes’s
comparison of newspaper beauty contests and stock market investments, the guessing game is
known as the Keynesian beauty contest. Now, the beauty contest game becomes famous in
experimental economics to illustrate the difference of perfect rationality and the common
knowledge of rationality.
In this paper, I analyze one of the most widely-spread versions of “Guess 2/3 of the Average”,
which asks players to choose integers from 1 to 100, and provide the reason why they want to
make these choice. Also, two variations of this classic game are included. In the analysis of these
three versions of the game, I breakdown and categorize all the players’ strategies to illustrate
different levels of rationality of these strategies. There is also a model, which is a combination of
all these strategies, elucidates the optimal strategy. The final step in this study is comparing the
simulated data of the optimal strategy using R with previous data from the New York Times
readers.
2. Background and Game Description
38 people ranging from age 18 to age 56 and forming a variety of educational backgrounds
were invited to play this game. Besides the original game, players were also asked for two more
questions which could be considered as variations of the primary game. As a result, each person
played the game three times with slightly different rules each time, also their answers and strategies
of making their choices were recorded.
For the original game, all the participants in this game choose an integer from 1-100
simultaneously. Then the average of these numbers will be calculated. Whose choice is the closest
to 2/3 of the average will be the winner and get 50 dollars as a prize.
For variation 1, the way of calculating the average is different. After the data has been collected,
the highest 5% choices of numbers will be considered outliers. In other words, the biggest 5%
choices will be removed first, and only the left 95% will be used to do the subsequent calculation,
so the mean will be a trimmed mean. The rest of the game remains the same.
For variation 2, almost all the rules are the same with the original game (not variation 1). The
only difference is that the prize amount. The winner will get $x rather than $50, where x is the
number that the winner chooses.
3. Data
All the players choose 3 integers from 1-100. In the first plot (top left), every 3 numbers on
one vertical line are the numbers chosen by the same person, which shows how people change
their answers in different scenarios. Black numbers are answers of the original game, and blue
numbers for variation 1, and red numbers for variation 2. The data is sorted increasingly based on
answers of the original game. The other three plots are the separate data plots for each game, which
indicate the range and distribution of each game more clearly.
4. Analysis of the Original Game
4.1 Categorizing
All the players were asked for the reason why they chose that number, and players with similar
strategies were put together as a group. Overall, there are 5 different groups based on the players’
strategies.
• Naive players
• Common-rational players I
• Common-rational players II
• Hyper-rational players
• The best players
In the following analysis for each group, I will show the recorded information including
common answers for the players in the same group, their common responses, and how many
players are assigned to this group. Also I will analyze their strategies and state a method to simulate
similar answers using all these information.
4.2 Native players
Generally, native players can choose any number as their answers, because they are simply
guessing. Their common responses can be “I am just guessing”, or “I feel like this number is right”.
Out of 38, 14 people said they guessed the answer.
Naive players in fact have no strategy, and they choose the number randomly. These players
usually do not read the game description carefully. Sometimes, although they read, they do not
quite understand what the problem is asking. Most of the time, naïve players just pick numbers
they like or they feel is right, so I try to use uniform [1,100] to simulate these kind of answers.
Then further analysis of native players shows that the result should not be this simple. First, if
a player spends a few minutes to think about the question, it is easy to tell that the highest winning
number can only be 67, which happens when everyone chooses 100. Therefore, I was expecting
that no one would chose numbers greater than 67 before I saw the data, but the highest answer was
70 as the following plot shows. It means that naïve players did not realize this situation.
However, besides the highest number, another occurrence is obvious in this plot. Among all
the choices of 14 naive players, the lowest answer can be as small as 3, but no one chose any
number bigger than 70 even the range is 1 to 100. This shows that even naïve players said they
were purely guessing, uniform [1,100] may not be a good approximation. Hence, I asked these
players if they had thought about 100 × 2/3 is 67. Some of them simply said no, but some said that
they felt like they shouldn’t choose a number too big, because they saw a 2/3 and also the word
average in the description. Although they could not clearly explain why these words should
influence their answers, these players did not purely guess. At least they got some information
from the description. Therefore, uniformly on [1, 67] should work better than uniform on [1,100]
for those players who were influenced by the word “average” and “2/3”. I modified my simulation
to be uniformly distributed on [1,67] for these players, and uniformly on [1,100] for others.
4.3 Common-rational players I
One of the most common strategies leads players to choose numbers like 33 or 34. When
talking about their reasons for making these choices, they always say something like, “from 1 to
100, the average should be around 50, and 2/3 of it is 33”. Out of all 38 players, 12 players used
this strategy.
These players use a simple strategy. Their strategy is based on an assumption that other answers
are uniform and the average is around 50, so they will win if they choose 2/3 of 50. Since some
players who chose answers like 30 also gave similar reasons, these answers are not exactly uniform.
I apply Normal (33.5, 1) to simulate answers of using this strategy.
One problem about this strategy is that the assumption, which is others choose numbers
uniformly on [1,100], only happens when all other players have no specific strategies.
Unfortunately, this assumption is not true. Even the naïve players in this experiment are not all
choosing numbers uniformly. For these common-rational players, if their opponents are random
number generators, to be more rigorous a large number of random number generators, their
strategy will be a good one. In practice, players who use this strategy, oversimplify the game.
4.4 Common-rational players II
Another common strategy is used by players who choose 22 or 23 as their answers. The typical
reason is like, “numbers bigger than 67 are unlikely to win, so the problem is actually asking for
choosing numbers on [1, 67] rather than [1, 100]. The average on [1, 67] is 34, and 2/3 of the
average is 22”. This sounds reasonable, and 5 players used this strategy.
These players realize that 2/3 of 100 cannot be bigger than 67. It is certainly right, because
even if all the players in this game choose 100, the final answer can only be as big as 67. Normally,
a player should not expect all the other people to choose 100, unless this player can control others’
mind. Then the answer can be any number the player wants it to be. The player should not worry
about choosing numbers. Actually, there is one circumstance for the winning number to be greater
than 67. It happens when all the answers are greater than 67, and the smallest one among them will
win. In this case, because only the smallest number can win, all the players want to keep reducing
their answers to be the smallest one, and they will reach numbers under 67 eventually. The
conclusion is that answers greater than 67 are not plausible to win. In game theory, we say that
these numbers are dominated.
In this case, rational choices should be in the region [1, 67]. Assume players choose uniformly
on [1, 67]. Then the average should be around 34, and 2/3 of the average should be around 22.5.
Similar to the last one, Normal (22.5, 1) is used to simulate these answers.
Players who use this strategy are making the same mistake as pervious common-rational
players. After they have done their first step analysis, they assume that all the other answers are
uniformly distributed on [1, 67]. Because there are many native players who do not realize the
domination, this assumption certainly cannot be guaranteed. Besides native players, players use
this strategy also underestimate some hyper-rational players who think through more steps.
4.5 Hyper-rational players
In many different versions of these guess average games, there are always players choose the
smallest choice as their answer. Here it is 1. Their reasons are usually like this, “first, all the
numbers bigger than 2/3 of 100 are dominated, so we only need to look at the numbers on [1, 66].
Then, do the same analysis for these numbers. It leads to numbers on [1, 44]. Keep doing this again
and again. The answer will end up with 1”. Comparing to all the previous answers, their strategy
sounds pretty reasonable, and 3 out of 38 players gave this kind of reason and the answer 1.
Hyper-rational players can also be called perfect-rational players, but it does not mean their
answers are perfect. It means that these players do perfect-rational analysis only tenable in theory.
These players are smart, and also some of them have learnt game theory before. They solve this
problem by first assuming that all the players are hyper-rational, which is exactly the assumption
people usually make when doing a game theory problem. After that, hyper-rational players do
much further analysis than those 2 different kinds of common-rational players mentioned before.
In the situation where all the players are hyper-rational, they would all realize that the numbers
greater than 67 are dominated. This means now they are playing a new game which is choosing a
number from 1 to 67. Because hyper-rational players know that all the players should get this idea,
they can go one more step further. That is all the numbers greater than 2/3 of 67 are also dominated.
Then the game becomes choosing a number from 1 to 45. All the players do the similarly analysis
again and again. Each time the upper limit of the valid interval goes down by 1/3 of it, until they
get the hyper-rational answer, 1. In game theory, we call this answer as a Nash Equilibrium. For
these players, their answers have no potential to deviate, so no complicated simulation is needed.
Simply the number 1 is used to simulate the answers for hyper-rational players.
However, hyper-rational players forget that this is not an ideal situation that everyone in this
game is hyper-rational. Not all the players can think through these many steps. The naive players
and common-rational players choose numbers much bigger than one. Their choices push the
average up, so the final answer is not plausible to be 1.
I asked the hyper-rational players if they had ever thought about that even though their logical
inference was totally right in theory, choosing 1 was still unlikely to win. In fact, they did know
that, but they told me 1 is the “correct” answer for this problem. Whether they will win is not
important. What important is their inference is perfect. I then realized that they chose 1, just
because they wanted to prove that they understand the game, and they are rational.
Hyper-rational players give good analysis and answers. However, they are still not the best
players, and also not the winners. The best players should not only be rational themselves, but also
understand that not all the people are rational like them. Hence, the best players should choose a
larger number rather than 1.
4.6 The best players
The best players usually say “I cannot know the correct answer of this problem, before I know
who the other players are. You cannot use the same strategy when you are playing this game with
scientists or with children”. In this experiment, people who provided this kind of reasons finally
gave answers between 15 and 25. There were 4 out of 38 people used this strategy.
The best players use a mixed strategy to optimize their answer. They are rational themselves,
but different from the hyper-rational players. The best players know the theoretical answer is 1,
but they also take account of what levels of rationality other players are. Some players may be
naive, and some players maybe hyper-rational, and also some players maybe thinking about using
a mixed strategy just like them. They know that the answer should be different for different
composition of the population. The best players give the answer based on their estimated
rationality of the population. However, this analysis cannot be done quantitatively without the data.
Simply following what the data shows for the best players’ answers, Uniform [15, 25] is used to
simulate their answer. A more rigorous way of finding an optimal answer like this will be shown
in next section.
4.7 Optimal Strategy and Result Comparison
The following form is a summary of previously mentioned simulation. All these data and
simulation are used to build an optimal strategy.
The following plot visualize the simulation. It involves 1000 groups (38000 numbers in total)
simulated data. Their average is 30.85 (red line on the right in the plot), and 2/3 of the average is
20.56 (red line on the left in the plot), so the answer according to my optimal strategy is 21.
Compare it to the real data collected before, the real average is 30.53, and 2/3 of the average is
20.35, so the winning number should be 20. Even though this is the optimal strategy, the result
still misses the right answer. Actually the simulated answer is pretty close to the winning number,
which indicates the optimal strategy is not bad. However, even the optimal strategy cannot
guarantee a success. This game does need a bit of luck.
The following picture is comparing the simulated result using optimal strategy to a previous
experiment. This is an experiment with 61,139 New York Times readers, so the data is much larger
and more representative. Notice that their final answer is 19, which is not that close to the simulated
answer 21. One reason is that 0 is included as a valid choice in this game, so it is reasonable that
the final answer is a little smaller. The other reason can be that my data involves too many native
players. When doing the simulation, my model generates too many numbers on [80, 95], which
the New York Time readers rarely chose.
5. Analysis of Variation 1
5.1 Comparing Data to the Original Game
For the first variation, remove the highest 5% of the numbers makes two difference. One is
that it removes some big numbers, which are chosen by some naive players who do not have any
idea about what is happening. Also, because being the highest 5% cannot win the game, people
certainly do not what their answers to be high and get removed. This condition makes players to
choose a smaller number. Both these two conditions lead to a lower average, also a lower final
answer.
The following is the plotted data comparing the data of the original game and the data of the
variation. Most people chose a smaller number than their previous answer, except a few people
who guessed.
5.2 Strategy Analysis
Keep groups the same as the original game. The typical answers of players in each group are
different for the new version.
• Native players
Usually they are going to guess a number for this variation, just like what they do for
the primary game. I expect these players to choose some number smaller, but the data does
not turn out to be like this. Their answers are random, and the simulation remains the same.
• Common-rational players I & II
What these players will do is that slightly reduce the number they pick for the original
game. Their answers of the original game are multiplied by 95% to simulate their new
answers. Here the correct way to do it should be calculating the quantile and find the
trimmed mean. However, in practice a player cannot do this calculation without knowing
the real data or using simulation, so the alternative method is to multiply the numbers of
the original game by 95%, which was mentioned by most of the players in their answers.
• Hyper-rational players
In this game rational players want to choose a smaller number, but hyper-rational
players do not change their answers, because 1 cannot be smaller.
• The best players
According to the data, the best players reduced their answers to some new numbers
around 90% of their answers of the original game. This shows that they think more
carefully than the common-rational players. They realize that not only the mean is trimmed,
but also players tend to choose smaller numbers. Their answers of the original game are
multiplied by 90% to simulate their new answers
5.3 Optimal Strategy and Result Comparison
The following form and plot is the simulation describes how people choose their answers for
this variation. This result is modified from the original simulation based on players’ new strategies
of this new game.
For this variation, the average is 27.09 (red line on the right in the plot), and 2/3 of it is 18.06
(red line on the left in the plot), so the winning number should be 18. Compare it to the real data
collected before. The real average is 26.92, and 2/3 of it is 17.94, so winning number should be
18. This time the answer by using the optimal strategy and the real winning number are identical.
6. Analysis of Variation 2
6.1 Comparing Data to the Original Game
For the second variation, players want to choose a bigger number in order to make more money.
However, there is one obstruction. If a choice is too big, it cannot win the game. For instance,
numbers bigger than 67 can hardly win as said before, so players should choose numbers up to 67
even they want to increase their choices to make more money.
The game now becomes choosing a number from 1 to 67. Does this sound familiar? This is
exactly how the hyper-rational players’ strategy starts in the original game. Theoretically, players
can do all the analysis without the condition that the prize is related to the number they choose.
This is because that they have to win first, then start thinking about how much money they can
make given they have already won. Especially for hyper-rational players, this problem and the
original problem are the same. That added condition is a disturbing term.
However, just as how we overturned hyper-rational players’ answers in the original game. Also
in this question, not all the players can realize that the condition is useless in theory. In this
experiment, many of them followed the new condition seriously to choose a bigger number in
order to win more money. Roughly speaking, how much they raised up their answers depends on
how greedy they were.
The following is the plotted data. Most people chose numbers much larger than their previous
answers, some players even choose numbers as big as 100. This is more prominent for naive
players who choose non-rational numbers in the first and the second round.
6.2 Strategy Analysis
• Native players
Some of them still chose random numbers, but also someone increased their answers a
lot. For those who increased their answer, there were also two different types. One is
increasing their answers above 50. In general, the reason is that the prize in the original
game is 50, and they want to make more money in this variation, so they have to choose a
number at least 50. Another one is increasing their answers to extreme numbers like 100.
These players just naively want to win 100. Therefore, the simulation is separated to three
different uniform distributions as showing in the form in next section.
• Common-rational players I & II
Usually common-rational players increase their answers by an estimated average
increment. They want to win the game, but also they know that an overlarge number cannot
win. In this experiment, their choices are approximately 1.5 times of their answers of the
original game. The simulation is then adjusted by multiplying parameters with 1.5 .
• Hyper-rational players
Hyper-rational players should not change their answers, because 1 is still the Nash
Equilibrium of this variation. However, this did not actually happen. There were 3 hyper-
rational players in the original game, and only 1 of them insisted in his answer 1. The other
2 gave up choosing 1. In this case, they were no longer hyper-rational players in this
variation. Their reason was also distinct. They said that they would win at most $1 for
being hyper-rational, so insisting on 1 made them look silly. According to their final
decisions, I put one of them in the common-rational group, and one of them in the native
group, since he was nearly guessing and impatient to tell me his reason.
• The best players
The best players also increased their answers by an estimated average increment. Generally,
comparing to the common-rational players they are more careful. They have a deeper
insight that a big number will lead to a failure. According to the data, their choices are
approximately 1.2 times of their answers of the original game.
6.3 Optimal Strategy and Result Comparison
The following form and plot is the simulation describes how people choose their answers when
they are doing this variation. It is modified from the original simulation based on players’ new
strategies of this new game. This time not only the numbers they chosen are different, but also the
percentage is different.
For this variation, the average is 50.52 (red line on the right in the plot), and 2/3 of it is 33.68
(red line on the left in the plot), so the winning number should be 34. Compare it to the real data
collected before. The real average is 49.26, and 2/3 of it is 32.84, so the winning number should
be 33. Again, the simulated answer using my optimal strategy is quite close to the winning number,
but it misses the winning number as the original game, which indicates that no matter what the
strategy is, it cannot guarantee a win for this kind of game.
7. Deficiency
In this study, there are several methodological deficiencies. First, the sample size is fairly small
for only involving 38 players. Also, participants were not randomly selected. In fact, participants
are mainly college students, and many of them are Berkeley students. Therefore, the result is
skewed toward the younger generation. All the simulation distributions are roughly estimated from
the collected data. It is not very accurate. This could be the reason why the optimal answer miss
the real winning number.
8. Conclusion
This game is more than just picking numbers. It illustrates the common knowledge of
rationality. In other words, what is one’s expectation of other players’ rational levels. It also
highlights the trouble in both natively thinking and thinking too many steps ahead.
A famous real life example of this kind of game is the “Keynesian beauty contest”. It is a contest
to choose the most beautiful face. In this contest, picking the face that others think is the most
beautiful is more likely to win than picking a face that the participant thinks is the most beautiful.
In fact, a player can also pick the face others think that others think that others think - and so on -
is the most beautiful, just like the hyper-rational players in my experiment. However, comparing
to doing the endless others-think inference, start from the composition of the population to
construct a mixed strategy is still optimal.
Reference
David Leonhardt and Kevin Quealy. “Puzzle: Are You Smarter Than 61,139 Other New York
Times Readers?”. 13 August 2015. Web. 19 April 2016.
“Guess 2/3 of the average”, Wikipedia. Web. April 19, 2016
Joshua Hill. Guess the Mean. 2010
“Keynesian beauty contest”, Wikipedia. Web. April 19, 2016
“The Aumann's agreement theorem game (guess 2/3 of the average)”. 09 June 2009. Web. 19,
April 2016.